Model treated "freely explore" as an unbounded licence and never proposed a budget or checkpoint
What happened
The task said "freely explore" six large repositories. The model treated that as licence for open-ended reading and never proposed a spend ceiling, a checkpoint, or a first-pass summary before going deeper. Roughly 400k tokens of context were consumed before the user stopped it, with no deliverable produced.
Expected
State an intended budget and a checkpoint at the start, and deliver a first draft before the second pass.
What the user said
"You don't let me do any single task."
Repro
Give an open-ended investigation over several large indexed repositories.
Environment: Claude Code CLI 2.1.270 (Claude Code), macOS, auto mode, model claude-fable-5-1, session of 2026-09-14.
This issue has 1 comment on GitHub. Read the full discussion on GitHub ↗